Introduction

Globally, clouds contribute to surface cooling through reflection of shortwave (SW) solar radiation1,2. This cooling generally overwhelms the warming effect that clouds have because of their ability to trap longwave (LW) radiation emitted from the ground and the lower levels of the atmosphere. In the Arctic, however, where the sun is low in the sky, or even absent during the polar night, the warming effects of clouds dominate3. In fact, low clouds have been shown to warm the Arctic for most of the year by both observational4,5,6 and modelling7 studies. This process has been suggested to play an important role in the Arctic amplification of winter surface warming8.

On August 31st 2014 an effusive volcanic eruption started in Holuhraun in the central highlands of Iceland. It ended on February 27th 2015 after six months of continuous activity9. Emissions mostly stayed between the surface and 3 km above ground level10. Copious amounts of sulfur dioxide (\(\hbox {SO}_2\)) were released into the troposphere during the eruption, with observational estimates of total emissions ranging from 6.7 Tg11 to 9.6 Tg10. For comparison, satellite based estimates of combined anthropogenic \(\hbox {SO}_2\) emissions from Europe and Russia amounted to ca. 6 Tg for the entire year 2014 (e.g., Figure 13 in Fioletov et al.12). Through reactions with atmospheric oxidants these volcanic \(\hbox {SO}_2\) emissions led to an increase in sulfate (\(\hbox {SO}_4^{2-}\)) aerosols, referred to as \(\hbox {SO}_4\) aerosols hereafter. In gas phase, the hydroxyl radical (OH) is the most important oxidant, whereas the aqueous reactions within cloud droplets, which account for the majority of the global sulfate production13, are dominated by hydrogen peroxide (\(\hbox {H}_2\)O\(_2\)) and ozone (\(\hbox {O}_3\)). Additionally, transition metals acting as catalysts aid in the cloud aqueous chemistry14,15. The sulfate aerosols from the Holuhraun eruption altered the clouds in the vicinity of the laterally spreading volcanic plume by acting as cloud condensation nuclei (CCN)16.

Due to the amount of emissions, the eruption site in the relatively unpolluted environment of the northern Atlantic, and detailed observations, the 2014–15 Holuhraun eruption served as an excellent natural experiment of aerosol-cloud interactions (ACI) and how changes in low level clouds affect the climate. Previous studies have taken advantage of this, but so far the focus has mostly been on the SW albedo effects over the North Atlantic when it comes to climate impacts (e.g.,16,17,18,19). Here we take a closer look at the Arctic and argue that there, the 2014–15 Holuhraun eruption led to surface winter warming through an increase in low and middle level clouds and subsequent trapping of LW radiation.

Arctic surface warming

The fall of 2014 was unusually warm over the Greenland Sea. This is visible both in the ERA5 reanalysis20, see Fig. 1a and Supplementary Fig. S2, and the observational record, see Fig. 1b and Supplementary Figs. S1 and S3a. This warm anomaly is especially pronounced along the eastern coast of Greenland and over the southern Greenland Sea, reaching up to \(+\,\,3\,\,^{\circ }\)C for the September through November mean (SON), when emissions from the eruption were the highest, compared to the 1984–2013 linear trend. When inspecting modes of climate variability in the North Atlantic, it becomes evident that the North Atlantic Oscillation (NAO) was in a positive phase in SON 2014 (Supplementary Fig. S3b). This led to advection of warm air from the south over the Greenland Sea21. However, even when only considering years in which the SON mean NAO index was positive, we find that the fall of 2014 was the warmest fall during the 1984–2015 period for three out of the five meteorological stations used in this study (see Grímsey, Jan Mayen, and Ittoqqortormiit in Fig. 1b). We therefore turn our attention to the most prominent natural event in the vicinity at the time, namely the 2014–15 Holuhraun eruption.

In order to identify the climate response to the eruption, we perform an ensemble of ten unique and independent pairs of simulations using the Community Earth System Model version 2.1.3 with the Community Atmosphere model version 6, CESM2(CAM6)22. For each pair, the setup of the two members is identical apart from the inclusion of emissions from the Holuhraun eruption in one of them. All model components are active and coupled, and the meteorology is free-running. For these free-running simulations we calculate the difference within each pair to isolate the impacts of the eruption. For more details, see the “Methods” section. Using the ten member mean from the free-running simulations, we model a significant surface warming in the Arctic as a result of the eruption (Fig. 1c). This reflects the anomalously high SON temperatures over the Greenland Sea in 2014 in the ERA5 reanalysis (Fig. 1a), implying a volcanic contribution to the observed warming signal.

Fig. 1
figure 1

Surface air temperature anomalies for the September through November mean (SON) from (a) the ERA5 reanalysis for the year 2014 relative to the 1984–2013 linear trend, and (b) five meteorological stations around the Greenland Sea relative to the 1984–2013 linear trends, only for years with positive SON mean NAO. Panel (c) shows the SON mean surface air temperature response to the modelled Holuhraun eruption from ten free-running, fully-coupled CESM2(CAM6) simulations (see the “Methods” section for a description of the anomaly calculations). Anomalies inside (a) the 95% prediction interval based on the 1984–2013 climatology and (c) the 95% confidence interval are indicated with grey dots. Blue contours in (a) and (c) represent the sea ice edge as defined by 15% sea ice cover and the red triangles mark the location of Holuhraun.

Full size image

Cloud response to the eruption

We propose that \(\hbox {SO}_2\) emissions from the Holuhraun eruption led to warming over the Greenland Sea through formation of \(\hbox {SO}_4\) aerosols and subsequent changes in middle and low level cloud properties.

In addition to the free-running coupled simulations discussed above, we perform a pair of atmosphere only CESM2(CAM6) simulations where horizontal winds are nudged toward the MERRA-2 reanalysis23. The Holuhraun eruption is included in one of those simulations but not the other. In these nudged simulations, sea ice and sea surface temperature are prescribed. Figure 2d shows the maximum daily \(\hbox {SO}_2\) column from September to November 2014 in the nudged eruption simulation. The distribution and mass burden of the volcanic sulfur closely correspond to the Infrared Atmospheric Sounding Interferometer (IASI) retrievals11,24 shown in Fig. 2a. This indicates that the model well captures the observed \(\hbox {SO}_2\) transport and hence the volcanic cloud.

Fig. 2
figure 2

Panels (a) and (d) show the maximum \(\hbox {SO}_2\) column during September to November (SON) 2014 from IASI retrievals11,24 and the nudged CESM2(CAM6) simulations respectively. Panels (b) and (c) show SON mean anomalies in 2014 relative to the 2000 to 2023 climatological mean from CERES SYN1deg retrievals25 for cloud droplet effective radius and liquid water path respectively. The CERES SYN1deg retrievals are for low-level clouds, below 700 hPa. Panels (e) and (f) show SON mean anomalies from the nudged CESM2(CAM6) simulations for cloud droplet effective radius and liquid water path respectively.

Full size image

Figure 2b shows satellite retrievals from the Clouds and the Earth’s Radiant Energy System (CERES) SYN1deg dataset25 of cloud droplet effective radius during SON 2014 relative to the 2000 to 2023 climatological mean for low level clouds. As reported in previous studies (e.g., 16,18) the CERES retrievals show widespread and significant decrease in cloud droplet size around Iceland during the Holuhraun eruption. This cloud droplet response is also captured in the nudged CESM2(CAM6) simulations (Fig. 2e). The modelled negative anomalies are most prominent over the Greenland and Norwegian seas as opposed to being more evenly distributed across all ocean areas around Iceland in the CERES retrievals. The reason for this pattern in the modelled droplet size response is the low background CCN concentration at these high-latitude areas in the CESM2(CAM6) simulations (Supplementary Fig. S5a), leading to a disproportionally strong relative CCN increase (Supplementary Fig. S5c) despite a relatively weak absolute CCN increase (Supplementary Fig. S5b). Additionally, there are uncertainties in the CERES data due to retrievals errors. For example, large solar zenith angles at high-latitudes during winter lead to high uncertainties, particularly for thin clouds over bright surfaces. Furthermore, when only infrared channels are available, for example during the polar night, retrievals of cloud microphysical properties become highly uncertain26.

The Holuhraun eruption has also been found to have led to cloud adjustments27, with aerosol perturbations leading to increased cloud cover19,28. Here, we further find anomalously large cloud liquid water path (LWP) over the ocean areas around Iceland in SON 2014 in the CERES SYN1deg satellite retrievals (Fig. 2c). These positive LWP anomalies are especially prominent over the Greenland and Norwegian seas where they are partially significant and reach up to more than + 20 \(\hbox {g m}^{-2}\). Furthermore, they largely coincide with the positive temperature anomalies from the ERA5 reanalysis discussed above (Fig. 1a). Our nudged simulations also show increased LWP following the start of the eruption (Fig. 2f). The magnitude of the LWP increase over the Greenland and Norwegian seas is slightly higher in our simulations compared to the CERES retrievals. The CERES retrievals show negative LWP anomalies over the Greenland ice sheet in SON 2014 but these anomalies are insignificant (see also discussion on CERES uncertainties in the previous paragraph).

In their study, Malavelle et al.18 found that CAM5, the predecessor to CAM6, showed an excessive LWP response to aerosol perturbations from the Holuhraun eruption. In CAM6, the cloud microphysics scheme has been updated from diagnostic29 to prognostic30, resulting in the LWP being less sensitive to the cloud droplet number concentration. Of particular interest to this study, this especially applies where the cloud droplet number concentration is low as is often the case in the Arctic (see for example Supplementary Fig. S7a). This, along with the CERES retrievals discussed above, supports the plausibility of the cloud response to the volcanic aerosols in our simulations.

Fig. 3
figure 3

Mean SON anomalies from the free-running CESM2(CAM6) simulations: (a) \(\hbox {SO}_4\) aerosol column burden, (b) vertically integrated cloud droplet number concentration, (c) vertically averaged cloud droplet effective radius, (d) cloud liquid water path, (e) downward shortwave radiative flux density at the surface, and (d) downward longwave radiative flux density at the surface. The dotted regions indicate insignificance at the 95% confidence level, the orange contours \(+\,\,1\,\,^{\circ }\)C surface air warming anomalies from the free-running CESM2(CAM6) simulations, and the blue contours the mean SON sea ice edge from the eruption runs (15% sea ice cover defines the sea ice edge).

Full size image

Until now, we have discussed the cloud response to the volcanic aerosol perturbations in the context of nudged atmosphere-only simulations. In order to better understand the full climate response we turn our attention back to the free-running coupled simulations. In the free-running simulations, the volcanic \(\hbox {SO}_4\) aerosols are mainly transported over the seas north and east of Iceland (Fig. 3a and Supplementary Fig. S4b). This is similar to the spatial distribution in the nudged simulations (Supplementary Fig. S4a). As the aerosols are effective as CCN, they stimulate the formation of more numerous (Fig. 3b) but smaller (Fig. 3c) cloud droplets. Smaller droplets slow the formation of raindrops, reducing or delaying precipitation which leads to an increased LWP (Fig. 3d). For the most part, precipitation does not change significantly in our free-running simulations (see Supplementary Figs. S6 and S7 and accompanying discussion in Supplementary Section S.3), indicating that the modelled LWP increase is mainly due to delayed precipitation. This interpretation is in line with previous studies31,32,33. Additionally, there is an increase in middle and low level cloud cover over parts of the Greenland Sea and the Greenland ice sheet in our free-running simulations (Supplementary Fig. S8). This increase in cloud cover is quite substantial, reaching up to +8 percentage points for low level clouds along the east coast of Greenland, but not as pronounced as the LWP response in terms of significance. This is in line with the relatively limited cloud cover anomalies seen in our nudged simulations (Supplementary Fig. S9) and the mostly insignificant response we find in CALIPSO-GOCCP satellite retrievals34 (Supplementary Fig. S10).

The cloud droplet number concentration (Fig. 3b) and effective radius (Fig. 3c) anomalies in the free-running ensemble mean mostly coincide. However, the largest LWP anomalies (Fig. 3d), which are initially triggered by smaller cloud droplets, are advected from the open ocean between Jan Mayen and Svalbard towards the sea ice along the east coast of Greenland through the dominating circulation patterns in the region (Supplementary Fig. S11). Meanwhile, the ensemble mean cloud droplet effective radius remains unchanged in this area of maximum LWP increase. Aerosol perturbations in clean clouds lead to increased cloud droplet number and decreased cloud droplet size under constant LWP as the available cloud water is distributed to more droplets35. However, the strong LWP increase along the east coast of Greenland compensates for the higher cloud droplet number, resulting in the cloud droplet size to remain at its unperturbed level. Additionally, the anomaly patterns in the free-running ensemble mean should not be expected to exactly co-located as the meteorology differs between the ensemble members. For example, the LWP is controlled by a number of factors in addition to the cloud droplet number concentration, including temperature and synoptics.

Polar night covers the high Arctic during winter and insolation is limited in other seasons and at lower latitudes in the Arctic due to the low elevation of the sun. This means that any cloud changes have minimal effects on surface SW radiative transfer in the Arctic fall (Fig. 3e) and that the net radiation is LW dominated (Fig. 3f). It is this increase in LW radiation trapping under limited sunlight which is responsible for the surface warming depicted in Fig. 1. The diagram in Fig. 4 summarizes this mechanism. Note that the sign and distribution of the downward radiative anomalies presented in Fig. 3 closely match the net radiative anomalies presented in Supplementary Fig. S12.

Fig. 4
figure 4

Illustration of the mechanism leading to surface warming through increased LW radiation trapping by clouds under limited sunlight. The colored writing reflects the colormaps used for the respective quantities in other figures presented in this study.

Full size image

So far, we have explained the modelled warming signal as a result of cloud adjustments to aerosol perturbations, mostly through increased LWP. However, in addition to the LWP, the cloud droplet size is an important control for the cloud LW optical depth, their emissivity, and hence their LW trapping abilities. In particular, this applies to clouds with low LWP. In those cases, the LW optical depth increases as the cloud droplet size decreases3. Therefore, smaller cloud droplets as a result of the volcanic aerosol perturbations most likely also contributed to the surface warming signal.

The aerosol-cloud interactions discussed above can further alter circulation patterns through their impacts on radiation and temperature. In our free-running simulations, we find a significant decrease in the SON mean sea level pressure over the open ocean areas east of Iceland (Supplementary Fig. S11). This results in a slightly shifted and deeper subpolar low, accompanied with increased advection of warm air into the Greenland and Norwegian seas from the south. Furthermore, a deeper subpolar low is associated with a more positive NAO index (e.g., 36), which was, in fact, the case in SON 2014 (see discussion earlier in this study and Supplementary Fig. S3b). Our free-running simulations, therefore, indicate that volcanically induced circulation changes might have contributed to the warming signal over the Greenland and Norwegian seas alongside the LW trapping effects discussed above.

Due to the low elevation of the volcanic \(\hbox {SO}_2\) emissions (Fig. 5a) and the resulting volcanic aerosols (Fig. 5b), mainly low level clouds are affected (Fig. 5c–e). Additionally, the Arctic warming is visible throughout most of the tropospheric column, decreasing with altitude. It is, however, only significant close to the surface (Figs. 5f and 1c), demonstrating the effects of the LW radiation trapping.

Fig. 5
figure 5

Vertical profiles of SON zonal mean anomalies between \(30\,\,^{\circ }\)W and \(30\,\,^{\circ }\)E from the free-running CESM2(CAM6) simulations for (a) \(\hbox {SO}_2\) gas, (b) \(\hbox {SO}_4\) aerosols, (c) cloud droplet number concentration, (d) cloud droplet effective radius, (e) liquid water content, and (f) air temperature. The dotted regions indicate insignificance at the 95% confidence level and the red bars the emission latitude and altitude interval.

Full size image

Amplified warming over sea ice

In our free-running simulations, the sea ice extent is reduced regionally when the eruption is included. This is most prominent in the Greenland and Barents seas where we model a delay in sea ice growth following the start of the eruption (Supplementary Fig. S13). For more details, see discussion in Supplementary Section S.7 and Supplementary Fig. S14.

The surface air temperature increase is amplified over sea ice and dampened over open ocean in our simulations (Fig. 1b). This is also visible in the ERA5 reanalysis (Fig. 1a) but to a lesser extent. There are two main reasons for this: The availability of background CCN and the different thermal properties between open ocean and sea ice.

Over open ocean, CCN mainly originate on site, mostly as a result of sea spray and biological activity, both of which are included in CESM2(CAM6)37, resulting in the atmosphere being relatively CCN rich38,39. Over sea ice on the other hand, the atmosphere is relatively CCN poor since it has to rely on long range transport (Supplementary Fig. S5d). The relative increase of CCN over sea ice during the eruption is, therefore, much greater (ca. 300 to 500% increase) than over open ocean (ca. 100 to 300% increase), even though the absolute CCN increase over sea ice is less than half the increase over open ocean in our area of interest (Supplementary Figs. S5e and S5f). Accordingly, aerosols stemming from the Holuhraun eruption disproportionally affect clouds over sea ice (Fig. 3d).

Whereas the specific heat capacity of open ocean is approximately 4000 J \(\hbox {kg}^{-1}\) \(\hbox {K}^{-1}\)40, the specific heat capacity of sea ice is only about 2000 J \(\hbox {kg}^{-1}\) \(\hbox {K}^{-1}\)41. Additionally, the mixed layer depth in the open Arctic Ocean during winter has a magnitude of tens of meters42 whereas sea ice thickness in the Fram Strait can be measured in meters43. As a result, the open ocean warms much slower than sea ice for the same amount of absorbed energy and also has an overall greater heat capacity, hence buffering the LW trapping effects (see Figs. 1a, 1c and 3f). Also contributing to the different temperature responses are different strengths in radiative feedbacks between sea ice and open ocean. The Planck feedback, for example, is stronger for colder surfaces than warmer. As a result, the relatively cold sea ice needs to warm more than the relatively warm open ocean in response to the same change in radiative forcing (e.g., 44,45).

Temporal variations

The emission strength from the Holuhraun eruption varied considerably over the course of its six months duration, with \(\hbox {SO}_2\) emissions being strongest in the beginning of the eruption and weakest toward its end (Fig. 6a). This temporal behaviour is reflected in Fig. 6b which shows the monthly mean \(\hbox {SO}_4\) aerosol column burden anomalies from both the nudged and the free-running CESM2(CAM6) simulations, averaged over the area north of \(60\,\,^{\circ }\)N and between \(30\,\,^{\circ }\)W and \(30\,\,^{\circ }\)E, excluding landmasses. The Arctic, including the Greenland Sea, is relatively unpolluted and frequently CCN limited (e.g., 46 and Supplementary Fig. S18). As a result, the temporal evolution of the cloud droplet number concentration anomalies (Fig. 6c) is very similar to that of the aerosol mass. The cloud droplet size and LWP anomalies (Fig. 6d and 6e) follow a similar pattern. The correlation between the timeseries of the \(\hbox {SO}_4\) aerosol anomalies on one hand and the cloud anomalies on the other is generally very high. For the nudged simulations, the correlation coefficients amount to \(+\,\,0.90,\) \(-\,\,0.96,\) and \(+\,\,0.87\) for the cloud droplet number concentration, cloud droplet effective radius, and LWP respectively. For the free-running simulations, the corresponding correlation coefficients amount to \(+\,\,0.91\), \(-\,\,0.81\), and \(+\,\,0.79\). See Supplementary Fig. S15 for more details.

The importance of the combined effects of increased cloud LWP (Fig. 6e) and limited sunlight (as evident by the small SW anomalies in Fig. 6f) is illustrated in Figs. 6g and 6h which show the temporal evolution of the downward surface LW radiation and surface air temperature anomalies respectively. Despite scarce sunlight over the Greenland Sea in December to February, there is no significant temperature response during those months in the free-running simulations. The reason is the reduced LWP response, which has been weakened as a result of decreased aerosol perturbations. The warming signal reaches its maximum in November and then remains insignificant for the rest of the eruption, even though December and January receive less sunlight. From December onward, the volcanic \(\hbox {SO}_2\) emissions have decreased considerably compared to the first three months of the eruption. This indicates that the emission rate is crucial when it comes to the eruption’s impact on clouds and subsequent warming anomalies. Note that the limited LW radiation and temperature responses in the nudged simulations are a result of prescribed sea ice and sea surface temperature (see also Supplementary Fig. S16).

Fig. 6
figure 6

Prescribed daily Holuhraun emissions used in the CESM2(CAM6) simulations47 (a), and timeseries showing monthly mean anomalies from the nudged (purple) and free-running (blue) CESM2(CAM6) simulations for the area north of \(60\,\,^{\circ }\)N and between \(30\,\,^{\circ }\)W and \(30\,\,^{\circ }\)E, excluding landmasses; (b) \(\hbox {SO}_4\) aerosol column burden, (c) vertically integrated cloud droplet number concentration, (d) vertically averaged cloud droplet effective radius, (e) cloud liquid water path, (f) downward shortwave radiative flux density at the surface, (g) downward longwave radiative flux density at the surface, and (h) surface air temperature. The blue shades indicate the 95% confidence intervals around the free-running ensemble means, open circles insignificant free-running anomalies, and filled circles significant ones. The red lines in (b)–(h) indicate the duration of the eruption. Note that (a) shows daily values but (b)–(h) monthly mean values. In (a), the ticks on the horizontal axis indicate the first day of each month.

Full size image

Discussion

One obstacle in the way of climate research the Arctic, including the Greenland Sea, is the lack of in-situ observations compared to other more populated and easily accessible regions in the world. As a result, we have relatively few long observational timeseries in our area of interest. Further complicating the use of direct observations is the highly regional climate response to effusive eruptions. In our case, we use temperature data from five different sites in and around the Greenland Sea. Of those, two are significantly warmer in SON 2014 compared to the 1984 to 2013 climatology, three if only positive NAO years are considered, and none are anomalously cold.

When comparing the surface air temperature anomalies in the ERA5 reanalysis to the mean from the free-running CESM2(CAM6) simulations (Figs. 1a and 1c), differences in their spatial patterns and magnitudes stand out. The different spatial patterns can be explained by different meteorologies. In ERA5 the meteorology corresponds to SON 2014 whereas in the model simulations it is the mean of ten different atmospheric states. The magnitude of the significant anomalies in ERA5 mostly range between +1 and \(+\,\,2\,\,^{\circ }\)C, compared to about + 0.5 to \(+\,\,1.5\,\,^{\circ }\)C in the CESM2(CAM6) mean. One reason for this is that averaging over a ten member ensemble smooths out the response. Another reason is that the model simulations isolate the impacts of the volcanic eruption whereas the warming signal in ERA5 has multiple different contributing factors. Note that aerosols from the 2014–15 Holuhraun eruption are not included in ERA5 and hence their climate impacts are not explicitly simulated48. Instead, the impacts of the eruption are implicitly included in the reanalysis of other fields, such as surface air temperature. As the SON 2014 temperature in ERA5 is significantly higher than the 1984 to 2013 climatology over a large portion of the southern Greenland Sea, it cannot be fully explained by natural variability alone. To summarize, while our simulations do not demonstrate that the Holuhraun eruption caused the observed warming over the Greenland Sea in the fall of 2014, they indicate that the eruption was an important contributor to the warm anomaly.

Conclusion

Anomalously high surface air temperatures were observed over the Greenland Sea in the fall of 2014. In this study we link these warm anomalies to the 2014-15 Holuhraun eruption using an Earth system model and observational evidence. The volcanic sulfate aerosols caused larger cloud liquid water path, leading to increased trapping of longwave radiation under limited sunlight and subsequent surface warming. In their modelling study, Zambri et al.49 found that the large 1783-84 Laki eruption (effusive with explosive phases) in Iceland resulted in positive top of the atmosphere effective radiative forcing during winter at high northern latitudes. They attributed this change in radiative forcing to increased trapping of LW radiation through changes in cloud properties but did not identify any temperature response. However, that low clouds warm the Arctic for most of the year is well known and in this current study we show for the first time, to our knowledge, how an effusive volcanic eruption amplifies this effect.

While the 2014–15 Holuhraun eruption lasted for half a year, there is ice core based evidence for extended eruptive episodes in Iceland during the Holocene which lasted decades50. Also known are Icelandic eruptions whose total \(\hbox {SO}_2\) emissions far surpassed those from Holuhraun. An example is the 1783-84 Laki eruption mentioned earlier which is estimated to have emitted around 122 Tg \(\hbox {SO}_2\) into the troposphere and the lower stratosphere over a period of eight months51. Given the long history of active volcanism in Iceland, similar eruptions will almost certainly happen again in the future (e.g., 52). Understanding how the Arctic will response to such eruptions is highly relevant, especially given how the Arctic has warmed disproportionally fast compared to the rest of the world under the ongoing anthropogenic climate change (Arctic amplification)53. Future powerful and long lasting effusive eruptions in Iceland might lead to an even greater surface warming and hasten the occurrence of an ice free Arctic. Marine cloud brightening through cloud seeding has been suggested as one of the methods to combat anthropogenic climate change54. However, our results indicate that cloud seeding in the Arctic might lead to the opposite of what is intended, namely surface warming.

Methods

Model

The fully coupled Community Earth System Model version 2.1.3 with the Community Atmosphere Model version 6, CESM2(CAM6)22, is used for the model simulations performed in this study. CESM2(CAM6) has 32 vertical levels, extending to an altitude of 2.26 hPa (corresponding to ca. 40 km). In this study, the horizontal resolution is set to \(0.9\,\,^{\circ }\) latitude by \(1.25\,\,^{\circ }\) longitude. For radiative calculations, CESM2(CAM6) uses the Rapid Radiative Transfer Model for General circulation models (RRTMG)55,56.

The formation and development of the volcanic aerosols, along with aerosol-cloud interactions, are represented by the five main schemes listed below. (1) For sulfur chemistry, the simplified scheme of Barth et al.57 is used. It includes both gas-phase and aqueous reactions. (2) The four mode version of the Modal Aerosol Model (MAM4)37,58 simulates the formation and development of the aerosols. (3) The parametrization described by Abdul-Razzak and Ghan59 is used for aerosol activation as cloud condensation nuclei. (4) For cloud microphysics, the prognostic second version of the Morrison-Gettelman scheme (MG2)30 is used. (5) Finally, turbulence from grid-scale to shallow convection is simulated by the Cloud Layers Unified By Binormals (CLUBB) scheme60.

Experimental setup

A transient control run is carried out for the model years 2005 to 2015, using the CMIP6 historical forcing61. Every year between 2005 and 2014 an eruption case is branched off from the control and run for ten months. These simulations are fully coupled and all model components are active. Daily emissions from the Holuhraun eruption last from August 31st to February 27th the following year. A total of 8.3 Tg \(\hbox {SO}_2\) gas is emitted over this period, well mixed between 0.95 and 3 km above sea level, following a petrological estimate by Thordarson and Hartley47, validated by Schmidt et al.62. The emission altitude interval remains constant for the duration of the eruption. This emission scenario is detailed in Table 1. We refer to simulations using this model set-up as free-running. Fig. 7 illustrates the experimental set-up of the free-running simulations. For further discussion on the experimental design, see Supplementary Figs. S17–S19, and accompanying discussion in Supplementary Section S.10.

Fig. 7
figure 7

A schematic illustrating the set-up of the free-running CESM2(CAM6) simulations.

Full size image

Additionally, we perform a pair of simulations where the horizontal winds are nudged towards the MERRA-2 reanalysis23 over the period from August 1st 2014 to June 1st 2015. We nudge the winds in the model eight times per day (corresponding to the temporal frequency of the MERRA-2 data) and use a nudging strength of 12.5% (corresponding to a nudging coefficient of 0.125 in the terminology of CESM2(CAM6)). This results in a relaxation time of 24 h. One of the nudged simulation includes the same Holuhraun emissions as described above while the other does not. As before, CMIP6 historical forcing is used. As opposed to the free-running simulations, where all model components are active and fully coupled, the nudged simulations only have active atmospheric and river runoff components. The other are either inactive (ocean waves), prescribed (ocean, sea ice, and land ice) or run in satellite phenology mode (land). We refer to simulations using this model set-up as nudged.

Table 1 The Holuhraun emission scenario used in this study, based on petrological estimates by Thordarson and Hartley47.
Full size table

Observations, the ERA5 reanalysis, and satellite retrievals

We analyse surface air temperature data from five meteorological stations around the Greenland Sea, two from the Danish Meteorological Institute63 (Ittoqqortoormiit and Danmarkshavn), two from the Norwegian Centre for Climate Services64 (Svalbard lufthavn and Jan Mayen) and one from the Icelandic Met Office65 (Grímsey), and from the ERA5 reanalysis20 (downloaded from 66). Our results contain modified observational and ERA5 data and the providers are not responsible for any use that has been made of it. For analysis of observed distribution of sulfur from the Holuhraun eruption we use Infrared Atmospheric Sounding Interferometer (IASI) satellite retrievals of the \(\hbox {SO}_2\) atmospheric column11,24. For analysis of observed cloud droplet effective radius and liquid water path during the eruption we use Clouds and the Earth’s Radiant Energy System (CERES) SYN1deg satellite retrievals25. For analysis of observed cloud cover (in the supplementary information), we use CALIPSO-GOCCP satellite retrievals34.

Data processing

For the free-running simulations, we calculate anomalies for a variable \(\hbox {Y}_{\textrm{free},i}\) for each model year i such that

$$\begin{aligned} \Delta \textrm{Y}_{\textrm{free},i} = \textrm{Y}_{\textrm{free,erupt},i} - \textrm{Y}_{\textrm{free,contr},i} \end{aligned}$$
(1)

This results in an ensemble of ten sets of anomalies, one for each year. The two simulations which are being compared match on all background conditions (such as initial meteorology, background emissions, greenhouse gas concentrations, etc.), and only differ on a single aspect, namely the inclusion of the Holuhraun emissions. This approach is termed a matched-pairs analysis. For significance estimates, we calculate the 95% confidence interval (CI) using a two-tailed t-test such that

$$\begin{aligned} \textrm{CI} = \mu \pm t^* \cdot \hat{\sigma } \end{aligned}$$
(2)

with \(\mu\) being the ensemble mean, \(t^*\) an appropriate value from the t-statistics, and \(\hat{\sigma } = \sigma /\sqrt{n}\) the standard error of the ensemble. Here \(\sigma\) is the standard deviation of the ensemble and \(n=10\) the number of ensemble members. For a variable from the nudged simulations, \(\hbox {Y}_\textrm{nudged}\), the anomalies are simply calculated as

$$\begin{aligned} \Delta \textrm{Y}_\textrm{nudged} = \textrm{Y}_\textrm{nudged,erupt} - \textrm{Y}_\textrm{nudged,contr} \end{aligned}$$
(3)

We account for a long term trend in the observational and ERA5 temperature timeseries by subtracting a linear fit based on the 1984 to 2013 climatology. SON anomalies for a variable \(\hbox {Y}_{\textrm{tseries},i}\) at year i are therefore calculated as

$$\begin{aligned} \Delta \textrm{Y}_{\textrm{tseries,}i} = \textrm{Y}_{\textrm{tseries,}i} - \textrm{Y}_{\textrm{tseries,climfit,}i} \end{aligned}$$
(4)

where \(\textrm{Y}_{\textrm{tseries,climfit,}i}\) corresponds to the 1984 to 2013 linear fit evaluated for the year i. To estimate the significance of the anomalies from the timeseries we calculate the 95% prediction interval (PI) such that

$$\begin{aligned} \textrm{PI} = \mu \pm t^* \cdot \sigma \end{aligned}$$
(5)

where \(\mu\) is the mean of a timeseries, \(\sigma\) the standard deviation, and \(t^*\) an appropriate value from the t-statistics. Both \(\mu\) and \(\sigma\) are calculated from the 1984 to 2013 climatology. Again we use a two-tailed t-test. The data from the CERES retrievals is processed in the same way as the ERA5 data but using the 2000 to 2023 climatological mean as a reference when calculating anomalies.

Computational resources

The simulations were performed on the Fram high performance computer and the model output was stored in the National Infrastructure for Research Data (NIRD) storage system. Both Fram and the NIRD storage system are provided by Sigma2 and the Norwegian Research Infrastructure Services (NRIS) in Norway. We would further like to thank the anonymous reviewers for their helpful comments.